There are nine chapters in the book, as follows:

#1. An introduction to hypertext and cognition, Rouet et al.
#2. Studying and learning with hypertext: Empirical studies and their implications, Rouet & Levonen.
#3. Myths, misconceptions, and an alternative perspective on information usage and the electronic medium, Dillon.
#4. Using hypertext to study and reason about historical evidence, Britt, Rouet, & Perfetti.
#5. Effects of overview structure on study strategies and text representations for instructional hypertext, Dee-Lucas.
#6. Comprehension, coherence, and strategies in hypertext and linear text, Foltz.
#7. Studying and annotating electronic text, van Oostendorp.
#8. Notes on hypertext, cognition, and language, Espéret
#9. Text and hypertext, Perfetti.

I indicate chapters below by their numbers, e.g., #4.

Despite a number of influential books and articles (most notably by literary scholars), these authors claim that there is no general theory of hypertext (#2, p. 12): "ideology has often been substituted for scientific investigation," they say, "and although the hypertext literature is crowded with strong claims, rigorous empirical demonstrations are harder to find" (#2, p. 16). Or, as Dillon puts it in his chapter, theoretical claims about hypertext are transferred too readily into design without empirical testing: "Rather than using the real world as a shaper of theory, contemporary hypertext thinking seems tied to concepts of association and non-linearity of access even as it distorts them" (#3, p. 27). Dillon echoes here, of course, the discourse about hypertext familiar from the writings of Landow, Lanham, Bolter, Moulthrop, and others, where hypertext is located within an associative model of the mind; this is contrasted to the print medium, supposed to constrain the mind by its linearity and monologism.

Although passing reference is made in the book to authors such as Landow, the higher postmodern realms of hypertext are dismissed or ignored in order to focus exclusively on the cognitive and pedagogical aspects of hypertext. In particular, the authors examine design issues in relation to theories of reading and learning: like Charney (in Selfe and Hilligoss) they ask how far comprehension of text in hypertext can be modeled by cognitive theories of discourse. But they also suggest that discourse theory alone will be insufficient: we also need an account of the different aims that readers bring to hypertext, and we have to know more about the differences between readers themselves -- how much background knowledge they bring to bear on a hypertext task, or how familiar they are with hypertext functions.

In this commentary I outline the two main contributions that this book makes to that task. First, I describe discourse theory, the main framework for research on hypertext, and some of the findings from the empirical studies reported in the book (chapters 4 through 8 contain detailed accounts of actual studies). Second I outline Dillon's four-part ergonomic perspective and point out its implications for reading collections of hypertext documents, including the present one. While readers unfamiliar with research design and statistical analysis may find parts of the chapters in this book hard to follow, the implications of the studies often go beyond the specific contexts in which they were carried out and are worth close examination.

Understanding how human beings make sense of natural language, as it occurs in (say) a newspaper article or a short story, is a branch of psycholinguistics known as discourse or text theory. In a recent overview of the field, Graesser and his colleagues describe the different components of discourse as they are currently understood:

Most discourse psychologists adopt van Dijk & Kintsch's (1983) distinctions among the surface code, the textbase, and the referential situation model. The surface code preserves the exact wording and syntax of clauses. Comprehenders normally retain the surface code of only the most recent clause unless aspects of this surface code have important repercussions on meaning. The textbase contains explicit text propositions in a stripped-down form that preserves meaning, but not the exact wording and syntax. The textbase also includes a small number of inferences that are needed to establish local text coherence. The situation model is the content or the microworld that the text is about. The situation model for a story refers to the people, spatial setting, actions, and events in the mental microworld. This microworld is constructed inferentially through interactions between the explicit text and background world knowledge.

In addition, discourse theorists also study two other levels: the communication level (the communicative context within which a text is encountered) and the text genre level such as narration, exposition, description, persuasion, etc. (Graesser et al, 1997, p. 167).

Examinations of response to hypertext reported in Hypertext and Cognition are primarily located within discourse theory of this kind. For example, in his chapter Foltz describes the comprehension process in terms of the influential model of Kintsch, which has been effective at analysing and predicting how readers respond to expository prose and simple narratives. Folz emphasizes the mechanisms by which readers create a coherent understanding of text, building from local details (microstructure) to a sense of text meaning as a whole, the "macrostructure or gist of the text" (#6, p. 115). The "textbase" in this model can be represented as a hierarchical structure of concepts. It has been shown that readers are more likely to remember those propositions that occur higher in this hierarchy. But readers also bring their own knowledge to bear on understanding a text. As Folz notes,

The macrostructure is the result of the reader's inferential processes, with readers forming hypotheses of the overall meaning of paragraphs, chapters, and whole books. In this manner, the macrostructure incorporates the reader's background knowledge with the text in forming these hypotheses. The resulting macrostructure of the text is similar to the textbase in that it is also represented as a hierarchical structure with higher level concepts represented at the top. (#6, p. 115)

A full account of discourse, then, requires us to model both text structure and the reader's expectations and background knowledge, and how the two interact. (It should also be noted that discourse theory of this kind has not been particularly effective at understanding literary response: see Miall & Kuiken, 1994.)

The empirical studies reported in Hypertext and Cognition suggest that a hypertext intended to facilitate students' learning will be most effective if it draws on the constructive processes modeled by discourse theory. It is this issue I will examine now in more detail. It should be pointed out first that, at the present time, hypertexts for learning often depend on converting into hypertext form texts originally read conventionally. This may not be a fair test of the hypertext medium: as Dillon notes, "Converting linear text into hypertext may decrease comprehensibility of the materials, especially because coherence breaks down at the local and global levels" (#4, p. 46). In the study I focus on here the texts were created specifically for hypertext format, but this was not the case for the studies reported in other chapters where materials were adapted from printed texts (#4, #6).

In her study (Chapter 5) Dee-Lucas studied students' learning when reading a small hypertext on electricity consisting of nine nodes of text. She asked how effective were different kinds of overview of the nodes contained in the hypertext. The titles of the nodes were given either as an alphabetical list or as a hierarchical overview which mirrored the conceptual relationships between the nodes (shown in the Figure). In a third version the text was presented in a conventional linear format with the titles as headings and subheadings dividing the text (readers also studied this text on a computer screen). Students were asked to read the text either for general understanding or to read it in preparation for writing a summary. The readers who had the two hypertext overviews performed significantly better on recalling the titles of the nodes than readers of the linear text. Readers of the hierarchical overview were also markedly better at recalling the location of a title in relation to other titles in the overview. The hierarchical version was judged easier to read by the students asked to read for general understanding.

In general, then, the hierarchical overview provided the best conditions for learning. The reasons for its effectiveness are perhaps obvious. Firstly, it models the hierarchical structure of the textbase by identifying subordinate and superordinate concepts and the relationships between them. Secondly, a graphical representation of this kind exploits our capacity to benefit from spatial imagery in both understanding and remembering. While obvious, however, these implications for the design of hypertext are significant. They show, firstly, that reading is not a straightforward linear process: the construction of a macrostructure requires a process of making inferences, forming predictions, and holding in mind structural relationships between concepts at several levels. Readers may actually reread to review, or skim to gain an overview, or use similar non-linear strategies as they carry out these basic processes, as Foltz points out (#6, p. 120).

Moreover, these are processes that may be endangered or disabled during the reading of a hypertext. The so-called "freedom" of the hypertext reader may be empowering and lead to genuine discovery and new learning, but in practice the freedom of the hypertext user often seems constrained by the design of the hypertext -- by how access to text nodes is provided, by the meaning that can be attributed to links, by the ease of navigation between different conceptual levels, and similar factors, none of which is usually a consideration for readers of conventional print. As Rouet and Levonen put it, for effective reading the user "has to possess a mental representation of how the information is organized" (#2, pp. 12-13). This may work in favour of the reader: as in Dee-Lucas's example, the reader may profit from the navigational demands imposed by hypertext if its design mirrors the structural relationships inherent in the text; if the hypertext provides misleading or inadequate structural cues, on the other hand, the reader's ability to construct a model of meaning will be impaired. Studies of readers' disorientation in hypertext found readers looping around text nodes or flipping more or less at random: these behaviours point to navigation problems, such as failing to recognize nodes that have already been read, or not knowing how to find information when it is needed (#2, p. 16).

Hypertext, which requires readers to be active in choosing their own pathways, thus appears to impose a greater cognitive demand on the reader, who must attend to the navigation cues as well as to the task of comprehension (#2, p. 17). Hypertext may benefit from appropriate structural or locational cues, relational overviews such as that of Dee-Lucas, or by increased coherence of information through better semantic linking. In addition, readers may perform more effectively with explicit attention to reading strategies through exposure or training -- research studies are cited on each of these aspects (#2, pp. 18-20). But there appear to be inherent limits to hypertext design, arising from the variability between readers. We cannot design a hypertext that would accommodate every reader's particular reading strategies. Since readers' representations of macrostructure will vary, depending on their background knowledge, motivation, and other influences, it seems unlikely that hypertext links could be provided that would meet each reader's specific needs while reading.

Graphical representations suggest one aspect of this problem. It touches on what I would term the film-of-the-book problem. If you have read a particular novel then go to see a movie adapation, it is rather likely that the imagery on screen will strike you as unsatisfactory. It cannot carry the same significance as the imagery that you constructed for yourself during reading. Similarly, the imagery that I develop to support my developing understanding of a complex text almost certainly differs from that of the author of the text. To realize this imagery in a hypertext through spatial diagrams, or similar devices, may undermine or disable my own image-making functions. For rather simple learning tasks, such as the one modeled by Dee-Lucas, this may be unimportant. The benefit of the elementary hierarchical diagram to the student reader is well shown by her study. For more complex learning, however, the constructive process involved in producing appropriate imagery is itself a critical part of the learning process.

This raises the more general issue, what it means to learn. The hypertext models described in three of the empirical chapters (#4-#6) were primarily concerned with learning in the domains of history, electricity, and economic theory, where basic information or argument was conveyed by the hypertext. The studies focused on how effectively student readers were able to comprehend, that is build a mental model of, the macrostructure of the texts concerned, and how far hypertext design facilitated this process. In much learning, however, the representation of knowledge or argument is not so readily modeled: knowledge is contested; readers must interrogate a text and form their own model of its meaning. For much learning, the construction of relationships between propositions is itself a critical part of the learning process. Hypertext linking that makes this process explicit turns the reading task into one of memorizing prestructured information rather than promoting an active process of construction.

For these reasons hypertext may promote learning that is narrow, specific, and well-focused -- and there may be many occasions on which such learning is appropriate. In its current state of development, however, it may not be suited to providing the general learning environment that its advocates claim. Here is another piece of the evidence to show why. Much of the learning that we do (and all the learning we do as young children) is incidental learning: that is, it occurs without conscious intention or effort while we are engaged in other activities. It can also occur during explicit learning: while focusing on specific information in a text we may be absorbing other information that will also be available to us later in memory. Such incidental learning depends on a number of factors, particularly the motives currently in place in the learner, that is, what they are primed to pay attention to. Thus it is hard to predict or control incidental learning, yet often the most important learning occurs in this form.

In a second study, Dee-Lucas developed two different versions of a hierarchical overview: a complex one in which 22 nodes were represented, and a simpler one with only 9 nodes (see the diagrams in #5, p. 89). The same text was available in the hypertext, but two versions were built in which the text was segmented to correspond with one or other overview (i.e., the text was segmented into either 22 or 9 nodes). The text gave an account of a field of physics concerning the buoyancy of objects in fluid. Student readers presented with one or other hypertext were asked either to locate information (to define three technical terms), or to solve a problem in buoyancy by searching the hypertext for appropriate principles.

She found that when the task involved locating specific information the more highly segmented text was superior; when the task was to solve a specific problem both versions were equally effective. This leads Dee-Lucas to suggest that in general "providing more structure in a hypertext facilitates the development of efficient and effective study strategies. Readers with the more segmented hypertext went directly to the most specific task-related units in studying the text, and were less likely to select irrelevant units than readers with the less segmented hypertext" (#5, p. 97). So she goes on to claim greater efficiency in using the highly segmented text: "because structure increases study efficiency, readers are less likely to encounter related nontarget content when using highly structured overviews" (#5, p. 104). This suggests one acceptable model of learning. However, Dee-Lucas also notes that when examining what students recalled she found that whereas the amount of information remembered was the same in both conditions, those reading the less segmented version recalled more information that was not relevant to the task (#5, pp. 94-6). In other words, the less-segmented version facilitated incidental learning.

Although this is a minor example, its implications for hypertext as a learning tool are important. The segmentation of hypertext comes at a price. The more highly segmented the text, the less likely it is that incidental learning will occur, thus the background knowledge, motives, and personal imagery of the reader that contribute to incidental learning will be less engaged. The personal meaning with which we endow everday, incidental learning will be less available to contextualize and enrich the specific learning task that has been set. In other words, the more highly-tuned our hypertexts for specific learning goals, the more detached and isolated the learning that results. Such learning is more likely to be brittle and soon forgotten.

Dillon's refreshing and well-informed chapter begins by questioning the more extravagant claims made on behalf of hypertext. But he is no Luddite, he remarks: his only sin is "to question the computer metaphor of mind and to seek some evidence of the technology's impact on education that meets minimal standards of scientific rationality" (#3, p. 26). He discusses the "myths" that surround much hypertext advocacy: the myth that the mind is associationist in how information is linked, so hypertext is more natural; the myth that print is linear, hence constraining (evidence against this is provided: #3, p. 30); the myth that merely offering rapid access to a large textbase of information will be beneficial to learning.

In the second half of the chapter Dillon shows how his own framework for the development of documents can be extended to hypertext. It includes four analytical components, known as TIMS for short (#3, pp. 35-6):

• the Tasks or needs of the reader;
• the Information model of the documents' contents;
• the Manipulation skills and facilities available to the reader; and
• the Standard cognitive and perceptual reading processes of all reading.

As Dillon summarizes it,

document usage or reading is not a matter of merely scanning words on a page or of acquiring or applying a representational model of the text's structure, but a product of both these activities in conjunction with manipulating the document or information space and defining and achieving goals (all within a certain context). (#3, p. 36)

Dillon's broader approach, which he locates within ergonomics (or the human factors discipline, as it is also known), encourages a more complex approach to the issue of hypertext. He notes that shifts in priority may occur during any particular reading encounter, as the reader's strategy changes in the light of incoming information, or as a problem in comprehension is encountered. Thus it provides a set of parameters within which to study readers' behaviour as they engage with hypertext. In addition, his model also alerts a hypertext designer to the issues around which a hypertext can be evaluated. It can be used to predict the kinds of problem that a reader may face when using the hypertext (#3, p. 37).

To provide a pertinent example, consider the use that you as a reader are making of the hypertext documents in the present web site, including this one. The TIMS framework suggests that you should assess the relevance, accessibility, and coherence of the documents in relation to the purpose for which you are reading. First, what is that purpose, and how well do the documents available to you meet the expectations that represent that purpose? Second, to what extent are you able to understand the structure of information available in each document? Can you gain an impression of the document's relevance and structure reasonably quickly? Third, how possible is it to access the documents and navigate to the information you are seeking? Are document titles, content lists, or the index adequate? Fourth, is the writing at a level that is appropriate to challenge and sustain your interest without being too simple or too complex? As a committed reader (looking to complete the course partly with the aid of these documents) your evaluation of the documents in the light of these questions would be essential in any redesign or restructuring by the author.

While Dillon's model offers a wider and more productive framework for the design and evaluation of hypertext than any other author in the book, it is also important to bear in mind the field it is intended for: information-delivery and study systems, particularly in education. Dillon explicitly distances himself from the wilder shores of hypertext theory. Thus the "breakdown" model of Moulthrop is explicitly opposed to the model of textual coherence basic to Dillon's work. In the continuum of hypertext phenomena textual experiences that go beyond the primarily information model of Dillon (and the other authors in this book) may be misrepresented by being placed within this framework. We still stand in need of a general theory of hypertext, one that shows the place of feelings, or narrative expectations, or the unpredictable effects of irony, fantasy, or play. Hypertext may be good for these experiences too, but it would be equally misleading to claim that this all that hypertext should be.